Wind power space-time prediction method and system based on graph neural network and expansion time convolution

Abstract

The invention provides a wind power space-time prediction method and system based on a graph neural network and expansion time convolution, and relates to the technical field, and the method comprises the steps: obtaining a training set which comprises wind power plant data and a corresponding real power value; preprocessing the wind power plant data to obtain preprocessed wind power plant data; inputting the preprocessed wind power plant data into a constructed wind power prediction network to obtain a predicted power value; constructing a total loss function according to the predicted power value and the real power value, and training the wind power prediction network to obtain a trained wind power prediction network; and inputting to-be-detected wind power plant data into the trained wind power prediction network to obtain a prediction power result of the to-be-detected wind power plant data. According to the method, the graph neural network is constructed to accurately capture the spatial dependence of the fan, and the time dynamic characteristics are extracted in combination with the expansion time convolution depth, so that the problems of insufficient modeling of complex space-time characteristics and poor prediction precision of a traditional wind power prediction method are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of renewable energy power generation and comprehensive consumption, in particular to a wind power spatio-temporal prediction method and system based on graph neural network and dilated temporal convolution. BACKGROUND

[0002] Wind power prediction is generally divided into four levels: ultra-short-term (minute level), short-term (hour to day level), medium-term (week to month level) and long-term (quarter to year level). Ultra-short-term prediction is mainly used for wind farm control and real-time scheduling, short-term prediction serves for power grid load balancing and economic scheduling, and medium and long-term prediction provides reference for wind power planning and market transaction. At present, the mainstream prediction methods include the following categories: (1) Statistical modeling method: typical methods include time series model (ARIMA, SARIMA), polynomial regression, Kalman filter, etc. This kind of method fits the future trend through historical power data, which can provide certain accuracy under stable wind conditions. However, its essence is linear modeling, which is difficult to capture nonlinear characteristics such as wind speed mutation, gust disturbance and wake interference, and its prediction ability under complex weather conditions is limited.

[0003] (2) Physical modeling method: based on fluid mechanics and aerodynamics principles to construct wind energy conversion model, such as numerical wind field modeling method based on Reynolds average Navier-Stokes (RANS) equation or large eddy simulation (LES). Typical tools include FLORIS (NREL) and SOWFA, etc. This kind of model has strong interpretability, but it depends on high-precision meteorological input and complex parameter calibration process, and the calculation cost is high, which is not suitable for real-time prediction and large-scale wind farm application.

[0004] (3) Machine learning method: In recent years, support vector regression (SVR), random forest (RF), gradient boosting tree (GBDT) and other methods have been widely used in wind power prediction, which can automatically learn the complex mapping relationship between input features and output power, and improve the prediction accuracy. However, these methods are mostly based on single machine modeling, ignoring the spatial interaction between wind turbines, and are sensitive to noise and missing data, which is difficult to maintain stable performance in large-scale wind farm conditions.

[0005] (4) Deep learning methods: Deep neural networks (DNN), convolutional neural networks (CNN), and long short-term memory networks (LSTM) have shown good nonlinear fitting capabilities in the field of wind power prediction in recent years. CNN can extract spatial features, and LSTM can capture temporal dependencies. The combination of the two has improved the prediction performance to a certain extent. However, these models are usually based on regular grid structures to construct input features, which cannot accurately express the non-uniform distribution and dynamic wake coupling relationship between wind turbines. In addition, the models have large computational cost, weak interpretability, and poor adaptability to changes in dynamic spatial relationships, which limits their application in offshore wind farms. Summary of the Invention

[0006] To overcome the shortcomings in spatial dependency modeling and insufficient extraction of temporal dynamic features mentioned above, this invention provides a spatiotemporal prediction method and system for wind power based on graph neural networks and dilated temporal convolution.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides a spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution, comprising: Obtain the training set, including wind farm data and its corresponding actual power values; The wind farm data is preprocessed to obtain preprocessed wind farm data; The preprocessed wind farm data is input into the constructed wind power prediction network to obtain the predicted power value; A total loss function is constructed based on the predicted power value and the actual power value, and the wind power prediction network is trained to obtain a trained wind power prediction network. The wind farm data to be detected is input into the trained wind power prediction network to obtain the predicted power result of the wind farm data.

[0008] Preferably, the wind farm data includes: wind turbine geographical location, wind speed, wind direction, nacelle temperature, ambient temperature, pitch angle, reactive power, and active power.

[0009] Preferably, the preprocessing includes: anomaly detection and removal, missing value imputation, and feature standardization and encoding.

[0010] Preferably, the preprocessed wind farm data is input into the constructed wind power prediction network to obtain the predicted power value, including: The preprocessed wind farm data is transformed into a graph structure, the weights of the graph structure are obtained, and the weights are concatenated into an adaptive adjacency matrix. The adaptive adjacency matrix is ​​input into a preset graph neural network layer to obtain the final spatial feature matrix; The final spatial feature matrix is input into a preset dilated temporal convolution layer to obtain a final temporal feature matrix; The final spatial feature matrix and the final temporal feature matrix are spliced and input into a preset full connection layer to obtain a predicted power value.

[0011] Preferably, the preprocessed wind farm data is converted into a graph structure, the weight of the graph structure is obtained, and the weight is spliced into an adaptive adjacency matrix, including: The wind farm is abstracted into a graph structure:

[0012] The node set represents each wind turbine node in the wind farm; the edge set represents the spatial connection relationship between wind turbines; each node corresponding input features include wind speed, wind direction, cabin temperature, ambient temperature, pitch angle, reactive power and active power; each edge reflects the spatial coupling effect between wind turbines i and wind turbines j , and the weight is the geographical distance and power correlation between two wind turbines; The weighted adjacency matrix A between wind turbines is composed of the spatial weight between wind turbine nodes, and the elements are calculated according to the following formula:

[0013] , wherein represents the geographical distance between wind turbine i and wind turbine j, represents the historical correlation coefficient between wind turbine power sequences, is the relative weight of adjusting spatial distance and power correlation; The adaptive adjacency matrix is obtained by splicing: .

[0014] Preferably, the adaptive adjacency matrix is input into a preset graph neural network layer to obtain a final spatial feature matrix, including: The adjacency matrix is normalized to obtain a normalized adjacency matrix:

[0015] D is a degree matrix; The normalized adjacency matrix is subjected to graph convolution operation, and in the graph neural network layer of the first l layer, the spatial features of each node are updated through neighborhood information aggregation, and the calculation formula is as follows:​​

[0016] wherein, represents the input feature matrix of the l-th layer, l is a learnable weight parameter, is a nonlinear activation function; After L-layer graph convolution operation, the final spatial feature matrix is obtained. .

[0017] Preferably, the final spatial feature matrix is input into a preset dilated temporal convolution layer to obtain a final temporal feature matrix, including: obtaining a time sequence of the final spatial feature matrix , setting a time window , and the corresponding time feature sequence is represented as , wherein each represents a time feature at time step t. Performing dilated temporal convolution operation on the time feature to obtain a feature output:

[0018] wherein, represents the feature output at time step t, is a weight parameter of the k-th convolution kernel, represents a dilated convolution operation, is a bias term, is a Relu activation function, is a convolution kernel length; After multi-layer dilated temporal convolution, a final temporal feature matrix is obtained:

[0019] wherein, Dilated TCN is dilated temporal convolution.

[0020] Preferably, after the final spatial feature matrix and the final temporal feature matrix are spliced, they are input into a preset fully connected layer to obtain a predicted power value, including: splicing the final temporal feature matrix and the final spatial feature matrix according to the feature dimension to obtain a comprehensive feature matrix :

[0021] wherein, represents a splicing operation; inputting the comprehensive feature matrix into the fully connected layer to obtain the predicted power value:

[0022] wherein, represents a predicted power value at t+1 time, is a weight parameter of the kth convolution kernel, represents a dilated convolution operation, is a nonlinear activation function.

[0023] Preferably, the total loss function is a mean square error loss, and its expression is as follows:

[0024] wherein, T is the number of samples, is a predicted power at t time, is a true power at t time.

[0025] The application also provides a wind power spatio-temporal prediction system based on a graph neural network and a dilated temporal convolution, comprising: a data acquisition module, configured to acquire a training set, including wind farm data and corresponding true power values thereof; a preprocessing module, configured to preprocess the wind farm data to obtain preprocessed wind farm data; a predicted power acquisition module, configured to input the preprocessed wind farm data into a constructed wind power prediction network to obtain a predicted power value; a network training module, configured to construct a total loss function according to the predicted power value and the true power value, train the wind power prediction network to obtain a trained wind power prediction network; a power detection module, configured to input to-be-detected wind farm data into the trained wind power prediction network to obtain a predicted power result of the to-be-detected wind farm data.

[0026] Compared with the prior art, the technical scheme of the application has the beneficial effects that: The application firstly acquires a training set, including wind farm data and corresponding true power values thereof; then preprocesses the wind farm data to obtain preprocessed wind farm data; next inputs the preprocessed wind farm data into a constructed wind power prediction network to obtain a predicted power value; constructs a total loss function according to the predicted power value and the true power value, trains the wind power prediction network to obtain a trained wind power prediction network; finally inputs to-be-detected wind farm data into the trained wind power prediction network to obtain a predicted power result of the to-be-detected wind farm data. The application solves the problems of insufficient modeling of complex spatio-temporal characteristics and poor prediction accuracy of traditional wind power prediction methods by constructing a graph neural network to accurately capture spatial dependence of wind turbines and combining a dilated temporal convolution to deeply extract time dynamic characteristics. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A flowchart of an embodiment 1 wind power spatiotemporal prediction method based on a graph neural network and dilated temporal convolution; Figure 2 A structural diagram of an embodiment 3 wind power spatiotemporal prediction system based on a graph neural network and dilated temporal convolution. DETAILED DESCRIPTION

[0028] The drawings are only used for illustrative purposes and cannot be understood as limiting the patent; In order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual product size; It is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0029] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0030] Embodiment 1 The present embodiment provides a wind power spatiotemporal prediction method based on a graph neural network and dilated temporal convolution, as shown in Figure 1 , comprising: Obtain a training set, including wind farm data and its corresponding real power value; Preprocess the wind farm data to obtain preprocessed wind farm data; Input the preprocessed wind farm data into the constructed wind power prediction network to obtain the predicted power value; According to the predicted power value and the real power value, a total loss function is constructed, and the wind power prediction network is trained to obtain a trained wind power prediction network; Input the to-be-detected wind farm data into the trained wind power prediction network to obtain the predicted power result of the to-be-detected wind farm data.

[0031] In the implementation process, first, a training set is acquired, including wind farm data and corresponding real power values thereof; then the wind farm data is preprocessed to obtain preprocessed wind farm data; then the preprocessed wind farm data is input into a constructed wind power prediction network to obtain a predicted power value; second, a total loss function is constructed according to the predicted power value and the real power value, the wind power prediction network is trained to obtain a trained wind power prediction network; finally, the wind farm data to be detected is input into the trained wind power prediction network to obtain a predicted power result of the wind farm data to be detected. The present application realizes joint modeling of the spatial topological relationship between wind turbines in the wind farm and the power time evolution law by fusing a graph neural network and a dilated temporal convolution model, and solves the problem of insufficient spatial dependence and time feature extraction of traditional methods.

[0032] Embodiment 2 The present embodiment provides a wind power spatio-temporal prediction method based on a graph neural network and a dilated temporal convolution, comprising: acquiring a training set, including wind farm data and corresponding real power values thereof; The wind farm data is preprocessed to obtain preprocessed wind farm data. The preprocessed wind farm data is input into a constructed wind power prediction network to obtain a predicted power value. A total loss function is constructed according to the predicted power value and the real power value, and the wind power prediction network is trained to obtain a trained wind power prediction network. The wind farm data to be detected is input into the trained wind power prediction network to obtain a predicted power result of the wind farm data to be detected.

[0033] The wind farm data includes wind turbine geographic location, wind speed, wind direction, cabin temperature, ambient temperature, pitch angle, reactive power and active power.

[0034] The preprocessing includes abnormal data detection and elimination, missing value filling, and feature standardization and coding.

[0035] It should be noted that in the present embodiment, the abnormal data detection and elimination: Z-score standardization is used to identify and propose abnormal data and data during wind turbine shutdown; the missing value filling: for the actual data, a moving average method is used for recovery to ensure the continuity of the time series; the feature standardization and coding: the continuous features such as power, wind speed and temperature are normalized by minimum-maximization, and the periodic variables such as wind direction and pitch angle are encoded by sine-cosine; data preprocessing can effectively improve data quality, reduce noise interference, and provide accurate and stable input data basis for subsequent spatial feature extraction and time feature modeling.

[0036] inputting the preprocessed wind farm data into a constructed wind power prediction network to obtain a predicted power value, comprising: converting the preprocessed wind farm data into a graph structure, obtaining weights of the graph structure, and splicing the weights into an adaptive adjacency matrix; inputting the adaptive adjacency matrix into a preset graph neural network layer to obtain a final spatial feature matrix; inputting the final spatial feature matrix into a preset dilated temporal convolution layer to obtain a final temporal feature matrix; splicing the final spatial feature matrix and the final temporal feature matrix, and inputting the spliced result into a preset full connection layer to obtain the predicted power value.

[0037] It should be noted that in the embodiment, the preprocessed wind farm data is converted into a graph structure, the weights of the graph structure are obtained, and the weights are spliced into an adaptive adjacency matrix, comprising: The wind farm is abstracted into a graph structure:

[0038] wherein, the node set represents each wind turbine node in the wind farm; the edge set represents the spatial connection relationship between wind turbines; each node corresponding input features include wind speed, wind direction, cabin temperature, ambient temperature, pitch angle, reactive power and active power; each edge reflects the spatial coupling effect between wind turbine i and wind turbine j , and the weight is the geographical distance and power correlation between the two wind turbines; The weighted adjacency matrix A between wind turbines is composed of the spatial weights between wind turbine nodes , and the elements are calculated according to the following formula:

[0039] wherein, represents the geographical distance between wind turbine i and wind turbine j, represents the historical correlation coefficient between wind turbine power sequences, is the relative weight of adjusting the spatial distance and power correlation; The adaptive adjacency matrix is obtained by splicing : ; The matrix reflects the spatial coupling relationship between all wind turbines in the wind farm.

[0040] The adaptive adjacency matrix is ​​input into a preset graph neural network layer to obtain the final spatial feature matrix, including: The adjacency matrix is ​​normalized to obtain the normalized adjacency matrix:

[0041] in, D It is a degree matrix; Perform graph convolution on the normalized adjacency matrix, and at the th l In the graph neural network layer, the spatial features of each node are updated by aggregating neighborhood information, and the calculation formula is as follows:

[0042] in, Indicates the first l The input feature matrix of the layer, These are learnable weight parameters. It is a non-linear activation function; After L layers of graph convolution operations, the final spatial feature matrix is ​​obtained. .

[0043] Through multi-layer graph convolution operations, wind turbine nodes can simultaneously obtain the spatial features of themselves and their neighboring nodes, enabling the modeling of spatial dependencies within the wind farm and the characterization of wake effects. The resulting spatial feature matrix... As the spatial representation of the wind farm at time step t, it is passed as input to the subsequent dilated temporal convolutional layer for temporal dynamic feature modeling.

[0044] It should be noted that, in this embodiment, the final spatial feature matrix is ​​input into a preset dilated temporal convolutional layer to obtain the final temporal feature matrix, including: Obtain the final spatial feature matrix Time series, setting time windows The corresponding time feature sequence is represented as Each of them This represents the temporal characteristics of time step t; Perform dilated temporal convolution on the temporal features to obtain the feature output:

[0045] in, This represents the feature output at time step t. The weight parameters of the k-th convolutional kernel are... This indicates the dilated convolution operation. For bias terms, For ReLU activation functions, is the length of the convolution kernel; After the multi-layer dilated time convolution, the final time feature matrix is obtained:

[0046] wherein, Dilated TCN is dilated time convolution; the time feature matrix characterizes the dynamic mode of the wind power changing with time, which will be fused with the spatial feature to realize the spatio-temporal joint prediction of the wind power.

[0047] It should be noted that in the embodiment, the final spatial feature matrix and the final time feature matrix are spliced and input into a preset fully connected layer to obtain a predicted power value, including: the final time feature matrix is spliced with the final spatial feature matrix according to the feature dimension to obtain a comprehensive feature matrix .

[0048] wherein, represents a splicing operation; the operation can jointly express the spatial dependence relationship of each wind turbine in the wind farm and the time evolution law, so as to obtain a more comprehensive power feature representation; the comprehensive feature matrix is input into the fully connected layer to obtain the predicted power value through linear mapping:

[0049] wherein, represents the predicted power value at t+1 time, is the weight parameter of the kth convolution kernel, represents a dilated convolution operation, is a nonlinear activation function.

[0050] It should be noted that in the embodiment, the total loss function is mean square error loss, and its expression is as follows:

[0051] wherein, T is the number of samples, is the predicted power at t time, is the real power at t time.

[0052] By minimizing the loss function L , the model parameters are optimized, and the network weights of each layer, so that the prediction result gradually approaches the real power output. After multiple rounds of training, the model can adaptively capture the spatial topological features and time-varying patterns inside the wind farm, and realize high-precision spatio-temporal joint prediction of offshore wind power.

[0053] Embodiment 3 The embodiment provides a system for implementing the method in Embodiment 1 or 2, as shown, comprising: Figure 2 A data acquisition module is configured to acquire a training set, including wind farm data and corresponding real power values thereof; A preprocessing module is configured to preprocess the wind farm data to obtain preprocessed wind farm data; A predicted power acquisition module is configured to input the preprocessed wind farm data into a constructed wind power prediction network to obtain a predicted power value; A network training module is configured to construct a total loss function according to the predicted power value and the real power value, train the wind power prediction network, and obtain a trained wind power prediction network; A power detection module is configured to input to-be-detected wind farm data into the trained wind power prediction network to obtain a predicted power result of the to-be-detected wind farm data.

[0054] The same or similar reference numerals correspond to the same or similar components; The terms used to describe the positional relationship in the drawings are only used for exemplary illustration, and should not be understood as a limitation on the patent; Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the implementation modes of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and impossible to exhaust all the implementation modes. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.​

Claims

1. A spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution, characterized in that, include: Obtain the training set, including wind farm data and its corresponding actual power values; The wind farm data is preprocessed to obtain preprocessed wind farm data; The preprocessed wind farm data is input into the constructed wind power prediction network to obtain the predicted power value; A total loss function is constructed based on the predicted power value and the actual power value, and the wind power prediction network is trained to obtain a trained wind power prediction network. The wind farm data to be detected is input into the trained wind power prediction network to obtain the predicted power result of the wind farm data.

2. The spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claim 1, characterized in that, The wind farm data includes: wind turbine location, wind speed, wind direction, nacelle temperature, ambient temperature, pitch angle, reactive power, and active power.

3. The spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claim 2, characterized in that, The preprocessing includes: anomaly detection and removal, missing value imputation, and feature standardization and encoding.

4. The spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claim 3, characterized in that, The preprocessed wind farm data is input into the constructed wind power prediction network to obtain predicted power values, including: The preprocessed wind farm data is transformed into a graph structure, the weights of the graph structure are obtained, and the weights are concatenated into an adaptive adjacency matrix. The adaptive adjacency matrix is ​​input into a preset graph neural network layer to obtain the final spatial feature matrix; The final spatial feature matrix is ​​input into a preset dilated temporal convolutional layer to obtain the final temporal feature matrix; The final spatial feature matrix and the final temporal feature matrix are concatenated and then input into a preset fully connected layer to obtain the predicted power value.

5. The spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claim 4, characterized in that, The preprocessed wind farm data is transformed into a graph structure, the weights of the graph structure are obtained, and the weights are concatenated into an adaptive adjacency matrix, including: Abstracting a wind farm into a graph structure: Among them, the node set Represents the individual wind turbine nodes in a wind farm; edge set This indicates the spatial connection between wind turbines; each node The corresponding input features include wind speed, wind direction, nacelle temperature, ambient temperature, pitch angle, reactive power, and active power; each edge Reflecting the fan i With wind turbine j Spatial coupling effect between them, weight The geographical distance and power correlation between the two wind turbines; The weighted adjacency matrix A between wind turbines is composed of the spatial weights between each wind turbine node. Its composition, and its elements are calculated according to the following formula: in, This represents the geographical distance between wind turbine i and wind turbine j. This represents the historical correlation coefficient between wind turbine power series. To adjust the relative weights of spatial distance and power correlation; use By concatenating the matrices, we obtain the adaptive adjacency matrix: .

6. The spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claim 4, characterized in that, The adaptive adjacency matrix is ​​input into a preset graph neural network layer to obtain the final spatial feature matrix, including: The adjacency matrix is ​​normalized to obtain the normalized adjacency matrix: in, D It is a degree matrix; Perform graph convolution on the normalized adjacency matrix, and at the th l In the graph neural network layer, the spatial features of each node are updated by aggregating neighborhood information, and the calculation formula is as follows: in, Indicates the first l The input feature matrix of the layer, These are learnable weight parameters. It is a non-linear activation function; After L layers of graph convolution operations, the final spatial feature matrix is ​​obtained. .

7. The spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claim 4, characterized in that, The final spatial feature matrix is ​​input into a preset dilated temporal convolutional layer to obtain the final temporal feature matrix, including: Obtain the final spatial feature matrix Time series, setting time windows The corresponding time feature sequence is represented as Each of them This represents the temporal characteristics of time step t; Perform dilated temporal convolution on the temporal features to obtain the feature output: in, This represents the feature output at time step t. The weight parameters of the k-th convolutional kernel are... This indicates the dilated convolution operation. For bias terms, For ReLU activation functions, The kernel length is [length]. After multiple layers of dilated temporal convolution, the final temporal feature matrix is ​​obtained: Dilated TCN is a dilated temporal convolution.

8. The spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claim 4, characterized in that, The final spatial feature matrix and the final temporal feature matrix are concatenated and then input into a preset fully connected layer to obtain the predicted power value, including: The final time feature matrix With the final spatial feature matrix By concatenating the features along their respective dimensions, a comprehensive feature matrix is ​​obtained. : in, Indicates a splicing operation; The comprehensive feature matrix is ​​input into the fully connected layer to obtain the predicted power value: in, This represents the predicted power value at time t+1. The weight parameters of the k-th convolutional kernel are... This indicates the dilated convolution operation. It is a non-linear activation function.

9. The spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claim 8, characterized in that, The total loss function is the mean squared error loss, and its expression is as follows: Where T is the number of samples, Let be the predicted power at time t. Let be the actual power at time t.

10. A spatiotemporal prediction system for wind power based on graph neural networks and dilated temporal convolution, used to implement the spatiotemporal prediction method for wind power based on graph neural networks and dilated temporal convolution as described in claims 1-9, characterized in that, include: The data acquisition module is used to acquire the training set, including wind farm data and its corresponding actual power values; The preprocessing module is used to preprocess the wind farm data to obtain preprocessed wind farm data. The predicted power acquisition module is used to input the preprocessed wind farm data into the constructed wind power prediction network to obtain the predicted power value; The network training module is used to construct a total loss function based on the predicted power value and the actual power value, and to train the wind power prediction network to obtain a trained wind power prediction network. The power detection module is used to input the wind farm data to be detected into the trained wind power prediction network to obtain the predicted power result of the wind farm data.

Images